The cost grows with the number of shards: every query makes one request per shard, and the slowest shard decides the latency. Queries that include the shard key go to one shard instead.